Stratified Sample¶
Sampling design — instantiates Representative Sampling Design
Partitions the population into meaningful strata, samples within each — often oversampling the small or high-variance ones — and reweights so no decision-relevant subgroup disappears from the estimate.
Stratified Sample divides the target into mutually exclusive strata chosen so that cases are similar within a stratum and differ across strata, draws a probability sample inside each one, and then reweights the results to reassemble an unbiased picture of the whole. Its defining move — the one thing true of it and false of every sibling — is disproportionate allocation paired with reweighting: deliberately over-representing a small or high-variance subgroup in the sample, then correcting for that tilt in the math, so a group too rare to survive a simple random draw becomes reliably estimable. It is a selection design, not a fielded protocol; it says how cases are partitioned and drawn, and leaves contact, response, and claim-wording to whatever mechanism carries it.
Example¶
A national statistics office estimates median household income, and must also report a defensible figure for a rural indigenous minority that is roughly 2% of households. A simple random sample of 25,000 households (illustrative) would land only a few hundred minority households — too thin for a stable subgroup number, and the aggregate would swamp them. The design instead partitions households by region crossed with community type. It allocates disproportionately, oversampling the minority stratum until it yields enough households for a reliable estimate, and samples randomly within each stratum. It then applies design weights — the inverse of each stratum's selection probability — so the oversampled stratum does not distort the national median.
The outcome is two credible numbers from one design: a solid national median and a defensible minority-specific median, neither achievable by an unstructured draw of the same size.
How it works¶
- Choose stratifying variables that are measurable on the frame and correlated with the outcome, so strata are internally homogeneous.
- Set allocation — proportional (mirror the population), equal (compare strata), or optimal/Neyman (weight by within-stratum variability and cost) — and oversample any subgroup too small for its own estimate.
- Draw within strata by probability selection.
- Weight back to the population with design weights, and check that each cell cleared its minimum size before trusting a subgroup number.
Tuning parameters¶
- Stratifying variables — more and finer strata sharpen subgroup precision but shrink each cell and multiply design complexity.
- Allocation rule — proportional versus optimal versus deliberate oversampling; trades national precision against subgroup precision for a fixed budget.
- Number of strata — resolution versus fragile, undersized cells.
- Weight trimming — capping extreme design weights cuts estimator variance at the cost of a small bias.
When it helps, and when it misleads¶
Its strength is guaranteeing subgroup visibility and, when strata are genuinely homogeneous, cutting variance for a fixed sample size — you learn more per dollar than a random draw does.
Its failure mode shows up when teams chase optimal allocation blind. Neyman allocation minimizes variance only if the within-stratum standard deviations are known and stable; guess them wrong and a "clever" allocation can do worse than plain proportional sampling.[1] Weights compound the trap: they can rebalance a subgroup you observed, but they cannot manufacture evidence for one you never sampled, and extreme weights inflate variance until a subgroup estimate is precise-looking and unstable. The classic misuse is post-stratifying on a variable that was never part of the design and treating it as if it had been engineered in. The guarding discipline is to stratify only on what the frame can actually measure, watch cell sizes, and keep weights bounded.
How it implements the components¶
stratification_scheme— the partition into strata and the allocation of sample across them.selection_method— the probability draw executed within each stratum.weighting_or_adjustment_plan— the design weights that reassemble an unbiased whole-population estimate.
It does not run a live response_pattern_monitor or bind a generalization_boundary to who responded — those belong to Representative Survey Protocol, the fielded protocol that a stratified design is usually plugged into.
Related¶
- Instantiates: Representative Sampling Design — the subgroup-protecting selection design at the archetype's core.
- Sibling mechanisms: Representative Survey Protocol · Audit Sample · Field Sampling Plan · User Research Panel · Quality Inspection Sample · Public Consultation Panel · Benchmark Dataset
Editorial Notes¶
Form Classification¶
Form family: Representation, Specification & Plan
Rationale: Stratified Sample operates as a static representation, map, specification, schema, or prospective plan that externalizes information because it partitions the population into meaningful strata, samples within each — often oversampling the small or high-variance ones — and reweights so no decision-relevant subgroup disappears from the estimate.
Independent corroboration: The frozen evidence defines Stratified Sample as 'Partitions the population into meaningful strata, samples within each — often oversampling the small or high-variance ones — and reweights so no decision-relevant subgroup disappears from the estimate', so its operative form is Representation, Specification & Plan.
Nearest alternative: Analysis, Modeling & Optimization — Stratified Sample includes features of an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution, but its defining operation is a static representation, map, specification, schema, or prospective plan that externalizes information.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Sampling and reweighting within population strata is canonical survey design.
Related originating lineages:
- Data Science & Analytics — Data science, analytics, and operational monitoring supplies a parallel or contributing lineage for the mechanism's defining operation: partitions the population into meaningful strata, samples within each — often oversampling the small or high-variance ones — and reweights so no decision-relevant subgroup….
- Mathematics — Mathematical modeling, proof, and abstract-structure practice supplies a parallel or contributing lineage for the mechanism's defining operation: partitions the population into meaningful strata, samples within each — often oversampling the small or high-variance ones — and reweights so no decision-relevant subgroup….
- Public Administration & Policy — Oversampling protects small groups.
Review resolution: The blind reviewers agree that statistics_experimental_design is the primary origin and differ only on alternate origin disagreement, domain reach disagreement. I preserve every independently explained alternate from both records rather than imposing a numeric cap. I retain single_lineage because the combined evidence shows one traceable formative lineage. The broader reach of multi_domain records portability separately from historical provenance; encyclopedia_synthesis=false preserves the affirmative synthesis judgment where either reviewer identified one.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
Stratified Sample is the reusable selection primitive several siblings sit on top of: a survey protocol draws its sample stratified, an audit stratifies its ledger by risk band, a quality inspection allocates across process strata. Keeping the partition-and-reweight logic in one mechanism means those consumers inherit subgroup protection without re-deriving it.
References¶
[1] Neyman allocation — the split of sample size across strata (Jerzy Neyman, 1934) that minimizes an estimator's variance for a fixed total size by allocating in proportion to each stratum's size times its standard deviation. It is optimal only when those standard deviations are known in advance; misjudged, it can underperform proportional allocation. withdrawn registry ↩